Date:	Thu, 2 Jun 2011 21:16:18 -0400
From:	Neil Horman <nhorman@...driver.com>
To:	Ben Hutchings <bhutchings@...arflare.com>
Cc:	netdev@...r.kernel.org, Jay Vosburgh <fubar@...ibm.com>,
	Andy Gospodarek <andy@...yhouse.net>,
	"David S. Miller" <davem@...emloft.net>
Subject: Re: [PATCH] bonding: reset queue mapping prior to transmission to
 physical device

On Thu, Jun 02, 2011 at 09:13:24PM +0100, Ben Hutchings wrote:
> On Thu, 2011-06-02 at 15:46 -0400, Neil Horman wrote:
> > On Thu, Jun 02, 2011 at 08:09:49PM +0100, Ben Hutchings wrote:
> > > On Thu, 2011-06-02 at 14:56 -0400, Neil Horman wrote:
> > > > On Thu, Jun 02, 2011 at 07:35:53PM +0100, Ben Hutchings wrote:
> > > > > On Thu, 2011-06-02 at 14:03 -0400, Neil Horman wrote:
> > > > > > The bonding driver is multiqueue enabled, in which each queue represents a slave
> > > > > > to enable optional steering of output frames to given slaves against the default
> > > > > > output policy.  However, it needs to reset the skb->queue_mapping prior to
> > > > > > queuing to the physical device or the physical slave (if it is multiqueue) could
> > > > > > wind up transmitting on an unintended tx queue (one that was reserved for
> > > > > > specific traffic classes for instance)
> > > [...]
> > > > > So far as I can see, this has no effect, because dev_queue_xmit() always
> > > > > sets queue_mapping (in dev_pick_tx()).
> > > > > 
> > > > it resets the queue mapping exactly as you would expect it to.  bonding is a
> > > > multiqueue enabled device and selects a potentially non-zero queue based on the
> > > > output of bond_select_queue.
> > > > 
> > > > > What is the problem you're seeing?
> > > > > 
> > > > The problem is exctly that.  dev_pick_tx() on the bond device sets the
> > > > queue_mapping as per the result of bond_select_queue (the ndo_select_queue
> > > > method for the bonding driver).  The implementation there is based on the use of
> > > > tc with bonding, so that output slaves can be selected for certain types of
> > > > traffic.  But when that mechanism is used, skb->queue_mapping is preserved when
> > > > the bonding driver queues the frame to the underlying slave.  This denies the
> > > > slave (if its also a multiqueue device) the opportunity to reselect the queue
> > > > properly, because of this call path:
> > > > 
> > > > bond_queue_xmit
> > > >  dev_queue_xmit(slave_dev)
> > > >   dev_pick_tx()
> > > >    skb_tx_hash()
> > > >     __skb_tx_hash()
> > > > 
> > > > __skb_tx_hash sees that skb_queue_recorded returns true, and assigns a hardware queue mapping
> > > > based on what the bonding driver chose using its own internal logic.  Since
> > > > bonding uses the multiqueue infrastructure to do slave output selection without
> > > > any regard for slave output queue selection, it seems to me we should really
> > > > reset the queue mapping to zero so the slave device can pick its own tx queue.
> > > 
> > > So you're effectively clearing the *RX queue* number (as this is before
> > > dev_pick_tx()) in order to influence TX queue selection.
> > > 
> > No, you're not seeing it properly.  In bonding (as with all stacked devices) we
> > make two passes through dev_pick_tx.
> 
> Only if the stacked device has its own software queues and schedulers;
> e.g. VLAN devices don't.
> 
Every net_device has a default fifo qdisc, and as such, each net_device when
called with dev_queue_xmit calls dev_pick_tx, and that includes vlans.  It just
so happens that vlans are a single queue device and come out of dev_pick_tx with
a queue_mapping of 0.  Thats all moot of course, since we're talking about
bonding.

> > 1) The first time we call dev_pick_tx is when the network stack calls
> > dev_queue_xmit on the bond device.  here we call ndo_select_queue, which calls
> > bond_select_queue.  This method tells the stack which queue to enqueue the skb
> > on for the bond device.  We can use tc's skbedit action to select a particular
> > queue on the bond device for various classes of traffic, and those queues
> > correspond to individual slave devices.
> > 
> > 2) the second time we call dev_pick_tx is when the bonding bond_queue_xmit
> > routine (which is the bonding drivers ndo_start_xmit method, called after the
> > frist dev_pick_tx), calls dev_queue_xmit, passing the slave net_device pointer).
> > In this case we do whatever the slave device has configured (either reset the
> > queue to zero, or call the slaves ndo_select queue method).
> > 
> > What I'm fixing is the fact that the bonding drivers queue_mappnig value is
> > 'leaking' down to the slave.
> 
> And this is because dev_queue_xmit() assumes that it's a record of the
> RX queue number.
> 
Yes, correct.  stacked devices provide the opportuninty for queue_mapping to no
longer be that rx queue number, but rather a tx queue mapping.  This is a
potentential problem with all stacked devices, but bonding is the only thing
showing it right now because its the only stacked device that can set a non-zero
queue_mapping value.

> The differing interpretation of queue_mapping for RX and TX is annoying
> and I think we should change the initial value of queue_mapping to -1.
> But that's a separate issue.
> 
Agreed.

> > Lets say, for example we're bonding two multiqueue devices, and have the bonding
> > driver configured to send all traffic to 192.168.1.1 over the first slave (which
> > we can accomplish using an appropriate tc rule on the bond device, setting
> > frames with that dest ip to have a skb->queue_mapping of 1).
> > 
> > In the above example, frames bound for 192.168.1.1 have skb->queue_mapping set
> > to 1 when they enter bond_queue_xmit.  bond_queue_xmit, calls
> > dev_queue_xmit(slave, skb), which calls dev_pick_tx(slave, skb).  If we (for
> > simplicity's sake) assume the slave has no ndo_select_queue method, dev_pick_tx
> > calls __skb_tx_hash to determine the output queue.  But the bonding driver
> > already set skb->queue_mapping to 1 (because it wanted this frame output on the
> > first slave, not because it wanted to transmit the frame on the slaves tx queue
> > 1).  Nevertheless, __skb_tx_hash() sees that skb_rx_queue_recorded returns true
> > here and as a result, calls skb_rx_get_queue, which returns 0.  Because of that
> > we wind up transmitting on hardware queue 0 all the time, which is not at all
> > what we want.  What we want is for the bonding driver to be able to use
> > queue_mapping for its own purposes, and then allow the physical device to make
> > its own output queue selection independently.
> [...]
> 
> I think I understand - the bonding device is effectively single-queue
> and shouldn't record an RX queue number.  But I think you should define
Well, no, its a multiqueue device, it just happens to treat slaves as queues,
rather than internal hardware objects.  But I suppose you're saying the same
thing that dave is, in that we should restore the origional queue_mapping,
rather than clearing it to preserve the input semantics when we reach the
physical device.  I can do that.

Neil
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists